Transcribes images of single systems of printed music to ABC notation with a vision-language model, which Mx.ABC turns into a score.
The model is rokot-omr-2b from Hugging Face by default, or the Hugging Face id or local folder named in the MX_TRANSCRIBER_MODEL environment variable.
It runs in Python inside this OS process, through Pythonx: the first run installs Python, PyTorch and transformers with uv, and downloads the model into the Hugging Face cache.
Summary
Functions
The ABC text of each system image, in order, as {:ok, [abc]} or {:error, message}.
The width in pixels to give each system, about 1400 as the model card of rokot-omr-2b asks.